DefinePK

Abstract

Undetected cracks in materials like concrete, asphalt, metals, and composites jeopardize structural integrity, posing safety and economic risks across infrastructure, aerospace, and automotive sectors. This study proposes a Crack-Aware CNN-ViT Hybrid model for real-time crack detection, integrating a Crack-Aware Attention Module (CAM) to emphasize crack geometry and a Crack Severity Annotation Framework to classify cracks by width, depth, and impact. Trained on a 60,000-image RGB dataset, augmented with conditional Generative Adversarial Networks for diverse materials and conditions, the model achieves 95.3% ± 0.2% accuracy, 94.2% ± 0.3% precision, 96.0% ± 0.2% recall, 95.1% ± 0.2% F1 score, and 90.5% ± 0.4% IoU at 32 fps, processing webcam feeds on an NVIDIA Jetson Orin Nano. Ablation studies, cross-dataset validation on SDNET2018 and CrackTree260, and a real-world bridge inspection demonstrate statistically significant improvements over YOLOv8 (by 5.1% accuracy) and Vision Transformers. Enabling automated, edge-based monitoring with timestamped crack storage, this scalable solution advances structural health monitoring, ensuring predictive maintenance and safety.